Seventeen post-January releases clear the size, license and GPQA filters

A broad web scan found 17 distinct model releases reported above 30B total parameters and carrying Apache 2.0, MIT, or modified-MIT terms. Kimi K2.6 leads this permissive-license subset at 90.5% on GPQA Diamond; custom or research-only licenses were excluded.

17
qualifying releases
90.5%
highest reported GPQA
31B–1.6T
total parameter range
3
permissive license families

Reported GPQA scores

GPQA Diamond where identified · %
GLM-5.2
91.2
Nex-N2-Pro
90.7
Kimi K2.6
90.5
Hy3
90.4
DeepSeek V4-Pro
90.1
Kimi K2.7 Code
89.6
NVIDIA GLM-5.2 NVFP4
89.39
Qwen3.5-397B-A17B
88.4
DeepSeek V4-Flash
88.1
GLM-5.1
86.2
GLM-5
86.0
Qwen3.6-35B-A3B
86.0
Darwin-31B-Opus
85.9
MiniMax M2.5
85.2
Gemma 4 31B
84.3
DeepSeek R2
84.0
Command A+
76.0

What stands out

1
GLM supplies three of the strongest entries. GLM-5, 5.1 and 5.2 progress from 86.0 to 91.2 while retaining MIT terms. z.ai
2
Sparse models dominate the largest sizes. DeepSeek V4-Pro reaches 1.6T total parameters, while Kimi K2.6 reaches 1T, with only a fraction active per token. huggingface.co
3
The threshold admits one near-boundary dense model. Gemma 4 31B is reported at 30.7B precise total parameters, just above 30B. huggingface.co

Qualifying releases

ModelTotal paramsReleaseLicenseReported GPQASource
Qwen3.5-397B-A17B397B2026-02-01Apache-2.088.4, GPQA Diamondhuggingface.co
GLM-5744B2026-02-11MIT86.0, GPQA Diamondhuggingface.co
MiniMax M2.5229B2026-02-12Modified MIT85.2%, GPQA Diamondchatforest.com
Gemma 4 31B30.7B2026-04-02Apache-2.084.3%, GPQA Diamondhuggingface.co
GLM-5.1754B2026-04-07MIT86.2%, GPQA Diamondz.ai
Qwen3.6-35B-A3B35B2026-04-16Apache-2.086.0%, GPQA Diamondhuggingface.co
Kimi K2.61,000B2026-04-20Modified MIT90.5%, GPQA Diamondhuggingface.co
DeepSeek V4-Pro1,600B2026-04-24MIT90.1%, GPQA Diamondhuggingface.co
DeepSeek V4-Flash284B2026-04-24MIT88.1, GPQA Diamondfireworks.ai
DeepSeek R2671B2026-05-07MIT84%, GPQAllmcheck.net
Darwin-31B-Opus31B2026-05-14Apache-2.085.9%, reported GPQAhuggingface.co
Command A+218B2026-05-20Apache-2.076%, reported GPQAhuggingface.co
Kimi K2.7 Code1,020B2026-06-01Modified MIT89.6%, GPQA Diamondfireworks.ai
Nex-N2-Pro397B2026-06-02Apache-2.090.7, GPQA Diamondnerdleveltech.com
GLM-5.2753B2026-06-16MIT91.2, GPQA Diamondz.ai
NVIDIA GLM-5.2 NVFP4753B2026-06-25MIT89.39, GPQA Diamondhuggingface.co
Hy3295B2026-07-06Apache-2.090.4, GPQA Diamondhuggingface.co
Method: web search across model cards, publisher pages and independent reports, run 22 August 2026. “Parameters” means total parameters, not active parameters; scores are publisher- or source-reported and are not normalized across evaluation harnesses. Exact-date and benchmark-variant gaps are preserved in the table wording.
Keenable · made with SELECT* · 5,703 pages in 3m 34s · Open the chatShare: XLinkedInReddit